OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom-designed AI accelerator — an "Intelligence Processor" built specifically for inference rather than training. Unlike a general-purpose GPU, Jalapeño is an application-specific chip (ASIC) tuned around OpenAI's own understanding of how its models actually serve traffic: kernels, memory movement, and scheduling, all shaped by years of running ChatGPT and Codex at scale.
The chip reached tape-out in roughly nine months, a pace OpenAI calls the fastest ASIC development cycle ever achieved at this scale — accelerated in part by OpenAI's own models helping engineers draft and optimize the design. Engineering samples are already running production workloads in the lab, including GPT-5.3-Codex-Spark, with Broadcom CEO Hock Tan citing early performance-per-watt gains of roughly 50% over standard GPUs.
Gigawatt-scale deployment with Microsoft and other data-center partners is targeted for late 2026, under the umbrella of the broader Broadcom partnership (codenamed Project Nexus) announced last October. OpenAI still leans on Nvidia, AMD, and Cerebras for training and overflow capacity — this is a bet on owning the serving layer, not abandoning suppliers.
Anthropic launched Claude Tag, a beta feature for Slack that drops a persistent, channel-level @Claude into team workspaces. Unlike a personal assistant, Tag is a shared identity: anyone in the channel can delegate to it, watch its work in progress, and pick up where a teammate left off without re-explaining the task.
Admins grant per-channel access to tools, data, and codebases, and the assistant accumulates context the longer it stays in a channel — Anthropic says 65% of its own product team's code now comes from internal use of the same pattern. An optional "ambient" mode lets Claude post unprompted: surfacing a thread that's gone quiet, flagging a decision, or announcing a completed deploy.
Claude Tag replaces the older Claude-in-Slack app, which retires August 3, with a 30-day migration window for admins. It's a direct shot at the "AI coworker" category Microsoft Copilot, Glean, and Salesforce are all chasing from different angles.
SpaceX signed a compute-leasing agreement with Reflection AI, an open-weight model startup founded by former Google DeepMind researchers, for access to Nvidia GB300 chips at SpaceX's Colossus 2 facility in Memphis. Reflection will pay $150 million a month starting July 1, totaling up to $6.3 billion if the contract runs its full term through 2029.
Colossus was originally built to train Musk's Grok chatbot under xAI, now part of SpaceX. As that internal workload has plateaued, SpaceX has repositioned the site as a multi-tenant compute platform, having already signed larger deals with Anthropic (~$45B) and Google (~$30B), plus an arrangement with Cursor, which SpaceX is separately acquiring.
For Reflection, the deal is partly a statement: the startup is positioning open-weight models as a hedge against the kind of access restrictions large closed labs can impose. Both sides can exit with 90 days' notice after an initial three-month period.
In a long-form conversation on building what Databricks calls the Agent Cloud, CTO Matei Zaharia and Chief Architect Reynold Xin laid out why they think frontier AI infrastructure has to stay open. Their new framework, Omnigent, is a "meta-harness" that sits above existing coding agents — including Claude Code, OpenAI's Codex, and Inflection's Pi — giving teams one place to enforce cost budgets, permissions, and collaboration without touching what's underneath.
Zaharia points to a recurring pattern across Databricks' own 5,000-engineer org: production agent setups almost always mix harnesses and models — a cheap worker model paired with a frontier "advisor," or several models splitting planning, search, and generation in one flow. No single vendor's harness can express that natively, which is the gap Omnigent is built to fill.
The framework shipped under Apache 2.0, a licensing choice Xin frames as deliberate: removing the usual enterprise procurement objection to adopting infrastructure-layer tooling, and inviting the community to extend it directly.
Google's Fitbit Air launched at $99, paired with the Gemini-grounded Google Health Coach bundled into Google Health Premium at $9.99/month or $99/year — free for Google AI Pro and Ultra subscribers. The screenless, week-long-battery tracker exists mainly to feed the coach: continuous heart rate, SpO2, and sleep-stage data that personalizes its fitness, sleep, and recovery guidance.
Google says the coach was built against an internal "SHARP" evaluation framework — safety, helpfulness, accuracy, relevance, personalization — developed with a clinical advisory panel, and explicitly will not feed Fitbit health data into Google Ads.
Early reviews are split: praise for comfort and battery life sits alongside reports of the AI coach making accuracy mistakes and occasional hallucinated advice, a reminder that consumer health AI is being shipped at consumer prices before it's fully proven.